entropic-core 4.0.1

Privacy-first entropy monitoring with active LLM intervention for multi-agent systems

Entropic Core

Entropic Core is a reliability addon for agentic systems: it instruments long-running (single- and multi-agent) workflows with entropy-derived signals and provides hooks for regulation (interventions, healing, consensus) plus evidence (metrics, dashboards, artifacts).

Core Value (1 sentence)

For developers building AI agents, Entropic Core adds observability + entropy-guided control so agent systems remain stable and auditable during long runs.

What We Sell

Verifiable reliability for agent systems: fewer measurable failures with reproducible evidence.

What It Guarantees (and What It Does Not)

Guarantees

• A consistent control loop you can integrate: measure -> diagnose -> intervene -> record.
• Structured evidence: metrics, logs, and optional dashboards.
• Configurable policies to reduce manual tuning and make behavior more predictable.

Does NOT Guarantee

• "Zero hallucinations" universally. It can only reduce measurable failure modes under defined tasks and verification.
• Factual truth without grounding. If the system has no tools/data or reference, claims may be unverifiable.
• Scientific proof of "physics of information" claims. Strong claims require formal methodology and peer review.

Installation

pip install entropic-core

Fast Adoption: 1 Install, 1 Command, 1 Report

pip install "entropic-core[evals]"
entropic-killer-demo --provider mock --model mock-v1 --out entropic_dx_runs --max-cases 20 --no-smokes

Open:

• entropic_dx_runs/.../report.html

Commercial Clarity

What You Get Today (OSS)

• Drop-in runtime reliability loop for agent calls.
• Reproducible A/B evals with verifiable checks.
• One shareable report per run (report.html) plus artifacts (run.json, JSONL, CSV).

Commercial Roadmap

• Pro self-serve: hosted run history, scheduled evaluations, private dashboards, alerts.
• Team controls: shared workspaces, role-based access, audit timeline.
• Enterprise later: private deployment, SSO/RBAC hardening, compliance bundles, support SLA.

Not Available Yet

• 24/7 support and guaranteed incident response.
• Full compliance packages (SOC2/ISO) and formal procurement process.
• Custom enterprise integrations and dedicated success team.

Optional Extras

pip install "entropic-core[openrouter]"
pip install "entropic-core[dash]"
pip install "entropic-core[otel]"
pip install "entropic-core[evals]"

Quickstart

import entropic_core

# Recommended (drop-in): start a runtime session with an explicit mode.
# This creates a per-run folder with events + an HTML runtime report.
with entropic_core.protect(mode="balanced", provider="openrouter", run_dir="entropic_runs") as sess:
    # client = sess.wrap(client, provider="openrouter")
    pass

# Alternative: create middleware + wrap a provider client in-place
# (see docs/getting-started-openrouter.md for OpenRouter).
# mw = entropic_core.create_middleware()
# entropic_core.wrap(client, provider="openrouter", middleware=mw)

# Legacy: zero-config import-hook protection (kept for compatibility).
# entropic_core.protect()

Docs and Examples

• Manual: docs/README.md
• Examples: examples/README.md
• Public benchmark: evals/README.md (includes standard_v2 + entropic-killer-demo)

API Stability

• Stable API surface (SemVer, intentionally small):

• entropic_core.protect(...) (import-hook mode or runtime session mode)

• entropic_core.unprotect()

• entropic_core.create_middleware()

• entropic_core.wrap(client)

• Experimental APIs:

• entropic_core.experimental (modules)

• entropic_core.experimental.api (legacy convenience symbols; may change in MINOR/PATCH)

See DEPRECATION_POLICY.md and CHANGELOG.md.